Claude as an AI Agent: Full Setup Tutorial

Most people use Claude the way they use a calculator — ask a question, get an answer. But Claude with MCP servers is something entirely different. You give it a goal, and it uses real tools — your file system, the web, a browser — to actually complete it. This tutorial shows you exactly how to get there.

What You're Setting Up

Claude Desktop is Anthropic's desktop app for macOS and Windows. By default, it's a very capable chatbot. But Anthropic built it to support the Model Context Protocol (MCP) — a standard that lets you connect Claude to external tools called MCP servers.

Once you connect MCP servers, Claude can read and write files on your computer, search the web, control a browser, query databases, and call APIs. That's the difference between "answer my question" and "go do the thing."

The setup takes about 20–30 minutes the first time. After that, adding new tools is a 2-minute edit to a config file.

Step 1: Install Claude Desktop

Go to claude.ai/download and download Claude Desktop for your operating system. Install it and sign in with your Anthropic account. A Claude Pro subscription is recommended for heavier agent use — you'll hit rate limits faster on the free tier when an agent is making many calls in a loop.

Once installed, open Claude Desktop and confirm you can have a basic conversation. That's your baseline — Claude as a chatbot. Now you're going to upgrade it.

Step 2: Install Node.js

Most MCP servers are distributed as Node.js packages. You'll need Node.js installed to run them. Download it from nodejs.org — the LTS version is the right choice. After installation, confirm it works:

node --version
npx --version

Both commands should return version numbers. If they do, you're ready for the next step.

Step 3: Configure Your First MCP Server

Claude Desktop reads MCP server configurations from a JSON file. Find yours:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

If the file doesn't exist, create it. Here's a config that adds the filesystem MCP server — one of the most useful:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/yourname/Documents"
      ]
    }
  }
}

Replace /Users/yourname/Documents with the directory you want Claude to have access to. Keep it specific — don't give it your entire root directory yet.

Step 4: Add a Web Search Server

Research is one of the most powerful agent capabilities. Add the Brave Search MCP server (you'll need a free Brave Search API key from brave.com/search/api/):

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/yourname/Documents"]
    },
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-brave-search"],
      "env": {
        "BRAVE_API_KEY": "your-brave-api-key-here"
      }
    }
  }
}

Save the file, then fully quit and relaunch Claude Desktop. When the app restarts, it will load your MCP servers. Look for a hammer icon in the Claude interface — that's confirmation that tools are available.

Step 5: Run Your First Agent Task

Now the fun part. Give Claude a multi-step goal that requires both tools. Here's a good first test:

"Search the web for the three most popular productivity frameworks for software developers in 2025. Summarize each one in 2–3 sentences, then write all three summaries to a file called productivity-research.md in my Documents folder."

Watch what happens. Claude will call the web search tool, read the results, reason about the best answers, then call the file write tool to save the output. You'll see each tool call logged in the interface. When it finishes, check your Documents folder — the file will be there.

That's an AI agent. You gave it a goal with two distinct actions, and it completed both without you doing anything after the initial prompt.

Step 6: Expand With More Servers

Now that the basics work, you can add more capability. Here's what the most useful MCP servers unlock for you:

• @modelcontextprotocol/server-github — Read and write to GitHub repos, create issues, review PRs
• @modelcontextprotocol/server-postgres — Query a PostgreSQL database in natural language
• @modelcontextprotocol/server-slack — Read and send Slack messages
• @modelcontextprotocol/server-google-maps — Location lookups and directions
• puppeteer — Full browser control for web automation

Each one follows the same config pattern. Add the server block to your JSON, restart Claude Desktop, and the tool becomes available. For the full installation walkthrough, see our guide on installing MCP servers in Claude.

Tips for Getting the Best Results

Give Claude a specific, bounded goal — not "help me with research" but "find 5 articles about X, summarize each, and save to a file." Specific end states help Claude know when it's done. And include any constraints upfront: "don't write anything to disk without confirming with me first" is a valid instruction that Claude will respect.

Start with read-only tasks before giving Claude write access. Once you're comfortable with how it behaves on read operations, you'll have better intuition for what to expect when it starts writing.

People Also Ask

Can Claude remember context between different agent sessions?

Not natively — each new Claude Desktop conversation starts fresh. But you can work around this by having Claude read a "memory file" at the start of each session — a text file you maintain that holds relevant context. Some third-party MCP servers also provide persistent memory functionality.

What's the difference between Claude Desktop and the Claude.ai web app for agent tasks?

The web app at claude.ai doesn't support MCP servers, so it can't access local files or run tools on your machine. Claude Desktop is specifically built for this — it's the gateway to true agent capabilities with Claude. If you're doing agent work, you need the Desktop app.

How do I know which MCP servers are safe to install?

Stick to the official Anthropic MCP servers (the @modelcontextprotocol namespace on npm) and servers from well-known organizations. Read the source code before installing anything from an unknown author — MCP servers run locally on your machine and have real file system access.

Troubleshooting Common Setup Issues

Tools don't appear after restart: Check your JSON syntax — a missing comma or bracket will prevent the config from loading. Use a JSON validator like jsonlint.com. Also confirm that Node.js and npx are in your system PATH.

File access is denied: Make sure the path you specified in the filesystem server config actually exists on your machine and that you have read/write permissions on it.

Claude refuses to use a tool: Sometimes Claude asks for permission before using certain tools. That's by design — it's checking in before taking an action. Just confirm in the chat and it will proceed.